Fully Automatic Segmentation For Breast Ultrasound Sequence Based On Visual Saliency

Breast cancer is the most common cancer among women, which has a serious impact on women’s physical and mental health. According to medical clinical experience, the treatment success rate will be enhanced if most of the breast cancer could be detected early. Due to non-radioactive, non-invasive and highly accurate identification of benign and malignant tumors, etc., ultrasound imaging method has already become an important method for breast cancer diagnosis. In order to assist doctors improving the accuracy and objectivity of ultrasonic detection of breast cancer and reducing the rate of misdiagnosis of malignant tumors, many computer-aided diagnosis(CAD) methods have been widely used in breast cancer diagnostic process.Segmentation is one of the most important tasks in CAD systems for breast ultrasound. The current breast ultrasound segmentation method has two drawbacks: first, effective automatic segmentation is lacking. Semi-automatic method is an interactive mode which needs to input the position of the suspicious tumor and background by radiologists. Usually, the inputs are points or curves which are used as the initial conditions for generating the ROIs. Due to the need of operator intervention, the method cannot meet the requirements of real-time analysis and batch processing. Most of the current automated methods use only part of the image, which could exclude some background tissue. Then the BUS images extract from video and only one frame is used when segmentation. In the actual disease diagnosis, doctors need to examine not only a single frame, but also with video-assisted diagnosis. So the single frame method doesn’t take advantage of the information sequence.To solve these problems, we proposed a fully automatic segmentation for breast ultrasound sequence based on visual saliency. The main contents include:1. This paper presents a saliency model for automated tumor detection in breast ultrasound images. We get the mammary layer first. Then anatomy cue is proposed which is a robust rule to describe the tumor. The contrast cue is used in BUS image because the breast tumor appears relatively darker. Finally, these two saliency cues are consolidated. Experiments show that our method can distinguish the normal tissues from the tumor region well.2. This paper presents a saliency model for automated tumor detection in BUS sequence. First, we extract candidate reconstruction bases via the single saliency detection. Then we employ two cues to reconstruct the saliency map, common reconstruction error and auxiliary reconstruction error. After that, global mammary layer location and breast tumor positioning is performed. Finally, we get the final saliency map. The method can handle BUS images that single saliency method can’t detect correctly.3. This paper presents a co-segmentation of breast ultrasound images based on multiple-domain knowledge. It combines spatial and frequency domain prior knowledge, and introduces the idea of co-segmentation to segment BUS sequence. First, the tumor poses, position and intensity distribution is modeled to constrain the segmentation in the spatial domain, and then the frequency domain via modeling the phase feature and the zero crossing feature. Finally, BUS sequence information is built based on co-segmentation. The experimental results show that the proposed method can handle BUS images with low contrast and hypoechoic well and can segment BUS accurately.